serps.io
Guides
Beginner·6 min read

Keyword research that starts from your own site

Start from the queries you already win, judge volume and difficulty honestly, and cluster terms into one page each before you write.

Start from the queries you already win

Open your search analytics and export every query that earned an impression over the last few months. This is the only keyword data about your site that is not a model estimate, and you already own it. Sort it by high impressions with low clicks: those are queries where you appear and nobody chooses you, usually because you sit at the bottom of page one or because the page that ranks does not answer the question. Then sort for positions between eight and twenty, the band where a rewrite or a few internal links can plausibly move you onto page one. Every research project should start here, because the pages already exist. The mistake is opening a keyword tool first and building a list of terms you have no presence for, which quietly converts a research task into a content production program you may not have the capacity to finish.

Build seeds from how customers actually talk

Seed terms should come from language people used, not language you invented. Pull from sales call notes, support tickets, the search box on your own site, review sites in your category, and the questions buried in your onboarding replies. You are looking for the nouns customers use for the problem, which are frequently not the nouns you use for the product. A team selling what they call workforce orchestration will find customers typing employee scheduling software, and the gap between those two phrases is worth more than any volume estimate in the account. Write seeds down verbatim, including the awkward ones. Then expand each seed inside a keyword tool rather than starting from whatever the tool suggests first, so the output stays anchored to demand you have observed. If every seed came from your own marketing site, you have built a list describing your positioning rather than your market. One useful exercise: ask three recent customers what they searched before they found you, then compare their words against the headings on your landing page. The mismatch is your first batch of seeds, and it usually costs nothing to fix.

Expand with the results page itself

The search engine will tell you what it considers related if you ask it. For each seed, collect the autocomplete suggestions, the People Also Ask questions, the related searches at the bottom, and the headings used by the pages currently ranking. That gives you the subtopics the engine already associates with the term, which is better evidence than a tool's related-keyword score because it reflects what is actually being served today. Do this by hand for your ten most important seeds before automating anything, since reading twenty results pages teaches you the shape of a market faster than any export will. Note which questions recur across several seeds, because a question that appears for five different terms is usually its own page. One caution: suggestions are personalized and localized, so check them in a clean session or from the market you are targeting, not from your logged-in browser.

Read volume as a range and a shape

Every public search volume figure is a modelled estimate, usually a twelve-month average, and treating it as exact leads to bad calls. Two terms both reported at twelve hundred a month can behave completely differently: one steady all year, one driven almost entirely by a spike in November. Look at the monthly trend, not the average, before committing to write anything. Watch for the rounding floors in most tools, where everything small is reported as the same low figure, which hides a long tail that frequently converts better than the head. Use volume to sort candidates into rough tiers instead of ranking them precisely. A term reported at ninety a month that describes someone ready to buy is worth more than one at 5,000 that describes someone who will never purchase, and no volume number will tell you which is which. That judgement stays yours.

Judge difficulty against your own site, not in the abstract

A difficulty score summarizes the link profiles of the current top ten and says nothing about you. The useful question is whether your site can compete with those specific pages. Look at the actual results: how many come from sites in your category versus large general publishers, how many are the page type you would build, and whether any of them are thin. A term with a middling score where every result is a major publisher is unwinnable for a young site. A term with a frightening score where three results are stale forum threads may be very winnable. Compare the referring domain counts of the ranking pages against what your best pages have earned, since that gap is the work. The failure mode is filtering a list to difficulty under thirty and writing whatever survives, which produces pages for terms nobody searches with any intent to act. Check how old the ranking pages are while you are there. A top ten where every result was published four years ago and nobody has updated since is a signal the incumbents have stopped defending the term, which matters more than any score.

Label intent before you judge the keyword

Sort every candidate by what the searcher wants: to learn something, to compare options, to reach a specific site, or to buy. Intent decides what you build, and getting it wrong wastes the entire page. If the top results for a term are all vendor comparison lists, a product page will not rank there however well written it is, because the engine has already decided the query wants a comparison. Check the live results before assigning intent rather than guessing from the wording, since terms that look commercial frequently return guides and the reverse happens just as often. Pay particular attention to the mixed results pages, showing two guides, two product pages, and a video, because those are opportunities: the engine is unsure, and a page serving the dominant reading can take the slot. Record the intent beside the keyword so whoever writes the brief does not have to re-derive it.

Cluster by shared results, then give each cluster one page

Two keywords belong on the same page when the same URLs rank for both. That is the only reliable test, and it is easy to run: pull the top ten for each term and group the ones that overlap by several results. Grouping by word similarity instead produces clusters the engine does not agree with, which is how you end up with two pages competing for one query and both ranking worse than a single page would have. Once clustered, pick the primary term for each group, usually the highest-volume one whose intent matches the page you want, and list the rest as secondary coverage. Then assign exactly one URL per cluster and write it down. If you already have two pages aimed at one cluster, decide now which one survives and redirect the other, because leaving both live is an ongoing cost that keyword research exists to prevent.

Write down the decision, not just the list

The output of keyword research is not a spreadsheet of terms, it is a short set of decisions someone can execute: this cluster, this URL, this intent, this page type, this is why we believe we can win, and this is what we will check in ninety days. Without the reasoning, the next person re-litigates every choice, and you cannot distinguish a bad bet from a good bet that has not paid out yet. Include the terms you rejected and why, because a rejection list stops the same unwinnable keyword being proposed every quarter. Set a review date. When it arrives, check whether the pages you shipped rank for the primary term, whether they picked up the secondary terms, and whether any of them ranked for something you never considered. That last category is the most valuable, since it is the market telling you what your research missed. Keep the document short enough that someone actually reads it. A page per cluster with six lines on it gets used, while a forty-tab workbook gets exported once and never opened again by anyone except the person who built it.